- Threatening or tipping a model generally has no significant effect on benchmark performance.
- Prompt variations can significantly affect performance on a per-question level. However, it is hard to know in advance whether a particular prompting approach will help or harm the LLM's ability to answer any particular question.
That 100% tracks expectation if your technical knowledge exceeds past “believer”.
Now… for fun. Look up “best prompting” or “the perfect prompt” on YouTube. Thousands of videos “tips” and “expect recommendations” that are bordering the arcane.
Haha, at least one this… I could make an excuse, that if I’m juggling prompts around, one that starts “You are a copy writer…” vs “You are an editor that…” lets me separate them with natural language vs some historically dubious file system disorganization.
The way I like to describe it is that you can't go from 1 developer to 0 thanks to AI, but you might be able to go from 10 to 9. Although not sure what the exact numbers are.
David Holz mentioned on Twitter that he was considering a Midjourney API. They're obviously providing it to Meta now, so it might become more broadly available after Midjourney becomes the default image gen for Meta products.
Midjourney wins on aesthetic for sure. Nothing else comes close. Midjourney images are just beautiful to behold.
David's ambition is to beat Google to building a world model you can play games in. He views the image and video business as a temporary intermediate to that end game.
It actually has impressive image generating ability, IMO. I think the two things go hand-in-hand. Its prompt adherence can be weaker than other models, though.
I think I'd push back and say that nerd culture is no longer really a single thing. Back in the star trek days, the nerd "community" was small enough that star trek could be a defining quality shared by the majority. Now the nerd community has grown, and there are too many people to have defining parts of the culture that are loved by the majority.
Eg if the nerd community had $x$ people in the star trek days, now there are more than $x$ nerds who like anime and more than $x$ nerds who dislike it. And the total size is much bigger than both.
I sometimes feel like we throw around the word fact too often. If I misspell a wrd, does that mean I have committed a factual inaccuracy? Since the wrd is explicitly spelled a certain way in the dictionary?
TLDR: That article is pretty low quality, and the "caused millions of people absolutely unrelated to it to have to ration water" doesn't seem like a reasonable conclusion. It's not mentioned at all in the source article. I took some notes on this article and traced back the research to the original article by The Austin Chronicle which is significantly better: https://www.austinchronicle.com/news/2025-07-25/texas-is-sti... , would recommend.
Main takeaways:
- Why are we building data centres so close to the equator where it's hot.
- It's depressing to see the high quality reporting from The Austin Chronicle watered down into more and more clickbaity soundbytes as it gets recycled through other "news" orgs. But at the same time, I wouldn't have heard about it otherwise.
- The water evaporation was interesting to me, and would love to read more on what percentage evaporates, and whether the Stargate plans to build non-evaporative cooling will actually hold out and how that'll impact the water grid.
- Would love some more info/context on that 463 mil number, but stopping my research here for now. Combining this with when/how often Texas has to ration its water would provide a stronger argument in support/against the provided claim of water rationing.
- The fact that we don't have good numbers for how much water data centres are using is crazy, we need that level of granularity/regulation.
- Markers of poor reporting:
- Numbers without context/clarity. Would it kill these sites to include a bar chart.
- Citations of sites that market engaging/entertaining
- Ambiguous / contradictory data
- Ambiguous references
Notes:
Interesting article! A few weird things:
1. The most cited reference is to a site called "Techie + Gamers", which self-describes itself as "TechieGamers.com is a leading destination for engaging entertainment coverage, news, net worths and TV shows with a strong focus on Shark Tank." Makes me suspicious of the journalistic quality of this and that article.
2. In the headline it says "Texas AI centers guzzle 463 million gallons". Further down it says "According to a July 2025 investigation by The Austin Chronicle, data centers across Central Texas, including Microsoft and US Army Corps facilities in San Antonio, used a combined 463 million gallons of water in 2023 and 2024 alone, as reported by Techie + Gamers." Over 2023 and 2024? Odd that it's giving the sum over 2 years. And not sure what it means that it includes the US Army? Also without any context I don't know what this number means.
- I checked the TechieGamers article and this contradicts what is written there, which says the 463 million number is for San Antonio alone.
3. Robert Mace, executive director of The Meadows Center for Water, notes that "once water evaporates, it's gone." This is interesting, not sure how much water is actually evaporated vs returned to the grid.
4. "The scale of water use is massive, as the Texas Water Development Board projections estimate that data centers in the state will consume 49 billion gallons of water in 2025, soaring to nearly 400 billion gallons by 2030, as per Techie + Gamers report. That’s about 7% of Texas’s total projected water use, according to the report."
- Mixed citations here, not sure whether these numbers are from Texas Water Development Board or Techie + Gamers. Also they project an increase from ~232 million gallons/year in 2024 to 49 billion in 2025? That's a 200x increase. And they expect a further ~8x increase from 2025 to 2030 to 400 billion? Or is it because the original number was only for Central Texas?
- 7% of what? The 2025 number or the 2030 number?
- Again subtle contradictions with TechieGamer which says "a white paper submitted to the Texas Water Development Board projected that data centers in the state will consume 49 billion gallons of water in 2025. That number is expected to rise to 399 billion gallons by 2030, nearly 7% of the state’s total projected water use.". So it's not the Texas Water Development Board but a whitepaper submitted to the board? Not sure who made these numbers now.
5. "Much of the water these centers use evaporates during cooling and can’t be recycled, a critical issue in an area already grappling with scarce water resources, as reported by Techie + Gamers."
- Again really want more info/numbers on this.
The root article seems to be from "The Austin Chronicle" :
1. This starts with "After Donald Trump and Elon Musk’s public breakup, Sam Altman replaced Musk as the president’s new favorite tech guy. Altman, the CEO of OpenAI, has become something like Musk’s archnemesis on the rapidly developing stage of artificial intelligence in Texas." This doesn't seem accurate with my reading of the news, and is so colourful that it makes me question the journalistic quality of this article.
2. The reporting across the three sources is mixed on who they're blaming. Economic Times doesn't even mention OpenAI and calls it "Microsoft's Stargate campus". Techi Gamers uses this phrase, but also later says "Microsoft has partnered with OpenAI". And The Austin Chronicle doesn't mention Microsoft at all and focuses on OpenAI. And the Wikipedia page for Stargate says "joint venture created by OpenAI, SoftBank, Oracle, and investment firm MGX." ?
3. I take it back reading it further this article is _significantly_ better than the others, with many more reputable sources.
4. Finally we get some real sources!! The 49 billion 2025 and 400 billion 2030 numbers are from HARC, Houston Advanced Research Center. And the 7% is actually 6.6%, and relative to the 2030 projection.
5. Finally real info on evaporation!! Still no numbers but we get a description of the process:
> Most data centers use an evaporative cooling system, in which the servers’ heat is absorbed by water. The heat is then removed from the water through evaporation, causing the water to be lost as vapor in the air. The cooler water then goes back through the machines, and this loop is regularly topped off with fresh water. After all, evaporation renders the water saltier and unusable after four or five cycles. “Then they dump the water, and it goes down the sewer,” Mace said.
> ...
> The Abilene Stargate campus will reportedly use a closed-loop, non-evaporative liquid cooling system that requires an initial refill of around 1 million gallons of water, with “minor” maintenance refills. Cook is skeptical that such closed-loop systems will use as little water as they suggest. It’s not possible, Cook says, to use the same water over and over again, recycled infinitely, to cool servers.
6. This article doesn't mention the 463 mil anywhere, which makes me think that was original research from TechiGamers. They reference SAWS, San Antonio Water System, but again the numbers are without context, so would need to do some original research to get any meaningful insights from these numbers.
The data is misleading at best, probably manipulated. Let’s not forget the data is from DB itself.
As an example, in 2018, the then boss of infra at DB, Ronald Pofalla introduced the “Pofalla Wende”, whereby a commuting train between two cities will simply turn around and not serve the rest of the line if it has a certain delay. In itself, one could argue it can make sense under certain circumstances. But all remaining stations on the line are not counted as delays, even if there is not replacement train.
Same with cancellations, these don’t show up as delays.
The situation is significantly more dire than what this data shows.
COVID harmed most industries that rely on manual labor. Suggested reasons vary, one is that people discovered they could do mental-work, and preferred it (and apparently it never occurred to them before). Another reason is many retirements and a gap in knowledge transfer to the next generation. Another is a huge increase in demand for non-physical work during COVID that never went away, leaving other industries short handed (and it's not a pay thing, there's not enough people for all demand in all industries, someone must run short).
Maybe the DB staff went soft during the lockdowns. Maybe some employees left and institutional knowledge suffered. The same thing happened with the truck driver shortage.
The way neoliberalism dealt with the public sector including rail service and infrastructure and then come up with "COVID killed Deutsche Bahn" is like saying that poor old sucker who was pushed down the staircase succumbed to his running nose. The problems run much deeper and were already visible in the 70s and 80s, but because it's only the public sector and rail traffic, not about more highways and more cars and then even more cars it never got fixed. Because who needs rails and trains, right.
data.whiteBall.v.x = data.whiteBall.v.y = data.blackBall.v.y = data.blackBall.v.x = 10;
reply